Intelligent defect detection method for CT (Computed Tomography) image of additive manufacturing multi-type lattice structural member

Abstract

An intelligent defect detection method for CT images of additive manufacturing multi-type lattice structural members belongs to the technical field of intelligent detection and image processing, solves the technical problems of post-processing delay, large calculation amount, low defect feature extraction efficiency and the like in the prior art, and comprises the following steps: S1, processing an original CT image by adopting a backbone network to obtain an original CT image; generating a plurality of feature maps with different scales; s2, integrating the feature maps of different scales generated in the step S1 by adopting a neck network, retaining shallow-layer space details and fusing deep-layer semantic information, and obtaining a feature map considering both positioning precision and semantic expression; and S3, comparing the feature map with the multi-scale information generated in the step S2 with the marked image data in the head network, calculating a loss function of the training set, and outputting position information, category and confidence of the defect in the image to complete intelligent defect detection. And high-precision and high-speed automatic detection and identification of defects of the lattice structural member are realized.

Description

Technical Field

[0001] This invention belongs to the field of intelligent detection and image processing technology, specifically relating to an intelligent defect detection method for CT images of additively manufactured multi-type dot matrix structural components. Background Technology

[0002] Metal additive manufacturing technology is widely used in industries such as aerospace, automotive, medical biotechnology, and energy because it can produce structural components with complex spatial geometries and excellent performance in a net-shape manner. Selective Laser Melting (SLM) technology uses a computer-controlled high-power laser beam to melt specific metal powder layer by layer to prepare metal components of the desired shape. This technology has significant advantages in manufacturing complex lattice structure metal components. However, lattice structures prepared by SLM technology inevitably contain macroscopic defects (such as porosity, lack of fusion, and column fractures), which reduces the service performance of the components.

[0003] To ensure the reliability of structural components in service, it is necessary to accurately detect macroscopic defects in the cells of the components. Industrial CT is a commonly used non-destructive testing method for detecting internal defects in lattice structures. Compared with other traditional non-destructive testing methods, CT technology is less affected by complex geometries and can achieve non-destructive imaging of the internal features of lattice structures.

[0004] In the existing technology, the defect detection technology for dot matrix structure components (additive manufacturing) based on industrial CT technology mainly focuses on statistical analysis of defects in metal additive manufacturing dot matrix structures using CT images, while there is little research on intelligent detection methods for automatic defect identification based on CT images. The patent document “An Adaptive Defect Detection Method, Device, Identification System and Storage Medium” (Publication No.: CN 113688807 A, Publication Date: November 23, 2021) improves Faster R-CNN by introducing an adaptive candidate frame module, thereby improving detection accuracy. Zhang et al. proposed an intelligent defect detection method for additive manufactured lattice structures based on a modified YOLOv3 model (Journal of Nondestructive Evaluation, 2022, 41(1)3), which can effectively identify typical internal defects in lattice substructure samples. Wen et al. proposed an improved YOLO v7 algorithm to identify FCC and BCC defects (Complex defects defection of 3D-Printed lattice structures: Accuracy and scale improvement in YOLO v7, IEEE Transactions on Instrumentation and Measurement, 2024, 73:1-9). The above technologies and methods suffer from problems such as severe post-processing delays, large computational loads, and low feature extraction efficiency.

[0005] Therefore, this invention proposes an intelligent defect detection method for CT images of additively manufactured multi-type dot matrix structural components, so as to achieve high-precision and high-speed automated detection and identification of macroscopic defects in metal additively manufactured dot matrix structural cells, and significantly improve its computational load. Summary of the Invention

[0006] The main objective of this invention is to overcome the shortcomings of existing technologies. Addressing the technical problems of post-processing delays, high computational loads, and low defect feature extraction efficiency in existing methods for detecting defects in metal additive manufacturing lattice structures using CT images, this invention proposes an intelligent defect detection method based on CT images of various types of additively manufactured lattice structures. This method automates defect detection based on a target recognition algorithm model, achieving high accuracy and automated identification of multiple defect types. Building upon the YOLO v10 target recognition algorithm, this invention introduces an attention mechanism and a Spatial Pyramid Pooling Deformable Fast (SPPDF) module to obtain more spatial details and global information about the target defects, enhancing the target's feature information and outputting a more expressive feature map, thereby forming a highly accurate intelligent detection model.

[0007] This invention is achieved through the following technical solution: an intelligent defect detection method for CT images of additively manufactured multi-type dot matrix structural components, comprising the following steps:

[0008] S1. A backbone network composed of several convolutional normalized activation functions and pooling functions with different parameters is used to process the original CT images of additive manufacturing multi-type lattice structure components, extract the features of the input images, and generate multiple feature maps of different scales.

[0009] S2. A path aggregation network consisting of upconvolution, upsampling and downsampling operations is used as the neck network to fuse the feature maps of different scales generated in step S1, thereby generating a feature map with multi-scale information and improving the detection accuracy of targets at different scales.

[0010] S3. In the head network, the feature map with multi-scale information generated by the neck network in step S2 is compared with the labeled image data, and the loss function of this training set is calculated. The loss function includes localization loss and classification loss. The location information, category and confidence of the defect in the image are output to complete the intelligent defect detection of CT images of additive manufacturing multi-type dot matrix structural parts.

[0011] Furthermore, step S1 involves processing the deepest feature map extracted from the backbone network using SPPDF and Pyramid Segmentation Attention (PSA). This leverages the adaptive receptive field adjustment capability of deformable convolution and the global representation learning capability of adaptive attention to obtain more effective semantic information about defective targets, thereby improving the model's detection performance. This includes the following steps:

[0012] S1-1. Acquire original CT images of various types of lattice structure components manufactured by additive manufacturing, adjust the adaptive sensitivity field through deformable convolutional layers, and simultaneously perceive regular or irregular shapes to improve feature representation, generate feature maps with rich spatial information, and obtain the original features.

[0013] S1-2. Perform three max pooling operations on the feature map obtained in step S1-1. The size of all pooling kernels is fixed at 5×5 and the stride size is 1. Extract global multi-scale information while keeping the feature size unchanged to obtain pooled features.

[0014] S1-3. The original features obtained in step S1-1 and the pooled features obtained in step S1-2 are concatenated along the channel dimension, and the channels are compressed through a convolutional layer to output the fused feature map. The feature map obtained after processing by the backbone network can not only retain more spatial and detailed information of the target defect, but also obtain further global information, thereby improving the robustness and detection accuracy of the present invention for multiple targets.

[0015] Furthermore, step S2 employs a neck network to introduce an improved Coordinate Attention (CALayer) module during the upsampling feature fusion stage. This module performs weighted feature mapping by encoding positions in different directions of the image space and performing a series of convolutional operations, outputting a more representative feature mapping. This includes the following steps:

[0016] S2-1. Encode the position of the input feature map along the width W and height H directions, concatenate the position codes with the feature mapping directions of the two adjacent layers, and feed them into different 1×1 convolutions respectively;

[0017] S2-2. Use a convolutional block consisting of the SiLU activation function, batch normalization, and convolution operation to obtain the channel dimension information C;

[0018] S2-3. In order to obtain better target feature information and output a feature map with stronger representation capabilities, the Sigmoid activation function and convolution operation are used to weight the feature map, retaining shallow spatial details and incorporating deep semantic information to obtain a feature map that balances localization accuracy and semantic expression.

[0019] Further, step S3 decodes the feature map output by the neck mesh in step S2, outputting the location information, category, and confidence level of the defect in the image, completing the intelligent defect detection of CT images of additively manufactured multi-type dot matrix structural parts, including the following steps:

[0020] S3-1. Use a multi-task joint optimization loss function to train the model:

[0021] L=λbox L CIoU +λ dfl L DFL +λ cls L BCE (1)

[0022] In equation (1), L CIoU The loss function for the Complete Intersection over Union (CIoU) bounding box; L DFL The distribution focal loss function; L BCE Let λ be the classification loss function; box , λ dfl , λ cls Both are weights of the subtask loss function, and λ box =7.5, λ dfl =1.5, λ cls =0.5;

[0023] S3-2. For each sample, define the true label g = [x g ,y g ,w g ,h g ,c g The prediction result is p = [x] p ,y p ,w p ,h p ,c p ], then L CIoU The loss function for a perfect intersection-union bounding box is:

[0024]

[0025] In equation (2), ρ 2 =(x g -x p ) 2 +(y g -y p ) 2 ;

[0026] IoU is the intersection-union ratio of the predicted bounding box and the ground truth bounding box; c is the diagonal length of the bounding rectangle of the two boxes;

[0027] S3-3. Let the boundary coordinates of the true label of the target to be detected be t. Discretize the possible values ​​of the boundary coordinates into 16 intervals, and use the network to predict the probability of each interval to obtain S. i and S i+1 L DFL The distributed focus loss function is:

[0028] L DFL =-((t) i+1 -t)log(S i )+(tt i )log(S i+1 (3) S3-4, Classification loss function L BCE for:

[0029] L BCE =-(g c log(p c )+(1-g c log(1-p) c (4)

[0030] In equation (4), g c p represents the label value for the category. c The confidence level of the class predicted by the network;

[0031] S3-5. The feature map is decoded through steps S3-1 to S3-4, and the location information, category and confidence level of the defect in the image are output.

[0032] The beneficial effects of this invention are as follows:

[0033] This invention, based on the YOLOv10 target recognition algorithm, introduces an attention mechanism and a fast deformation space pyramid pooling module to obtain more spatial details and global information about target defects, outputting more expressive feature maps. This forms a high-accuracy intelligent detection model, enabling high-precision, high-speed automated detection and identification of internal defects in additive manufacturing lattice structures. It solves the problems of existing technologies using CT images for defect detection in additive manufacturing lattice structures, which require human intervention for decision-making, and existing automatic defect identification methods suffer from post-processing delays, high computational loads, and low defect feature extraction efficiency. Attached Figure Description

[0034] Figure 1 This is a flowchart of the present invention;

[0035] Figure 2 Here is a flowchart of the SPPDF module;

[0036] Figure 3 Here is a flowchart of the CALayer module;

[0037] Figure 4 The following are photographs of the actual test specimens of the lattice structure in the embodiments; wherein, Figure (a) is the No. 1 body-centered cubic structure, Figure (b) is the No. 2 body-centered cubic structure, and Figure (c) is the diamond-like lattice structure.

[0038] Figure 5 This is a curve comparing the training loss values ​​of the present invention and the YOLOv10 model;

[0039] Figure 6 Figure 1 shows a comparison of the accuracy-recall curves of the present invention and the YOLOv10 model for each defect; Figure 2 shows a comparison of the accuracy-recall curves of the YOLOv10 model for each defect, and Figure 3 shows a comparison of the accuracy-recall curves of the present invention for each defect.

[0040] Figure 7 The figures show a comparison of the processing results of the present invention and the YOLOv10 model on CT images of dot-matrix structures; wherein, Figure (a) shows the processing result of the YOLOv10 model on CT images of dot-matrix structures, and Figure (b) shows the processing result of the present invention on CT images of dot-matrix structures.

[0041] Figure 8 Figure 1 shows a comparison of the CT slice processing results of the present invention and the YOLOv10 model on a dot matrix structure with prominent defects. Figure 2 shows the CT slice processing results of the YOLOv10 model on a dot matrix structure with prominent defects, and Figure 3 shows the CT slice processing results of the present invention on a dot matrix structure with prominent defects.

[0042] Figure 9 Figure 1 shows a comparison of the processing results of the present invention and the YOLOv10 model on CT slices of lattice structures with breakpoint defects. Figure 2 shows the processing results of the YOLOv10 model on CT slices of lattice structures with breakpoint defects, and Figure 3 shows the processing results of the present invention on CT slices of lattice structures with breakpoint defects. Detailed Implementation

[0043] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments.

[0044] This embodiment is used to identify internal defects in lattice structures of different sizes and morphologies (body-centered cubic and diamond-like structures). Two body-centered cubic structures (body-centered cubic component 1 and body-centered cubic component 2) of different sizes and one diamond-like lattice structure of 316L stainless steel components were fabricated using an HK-M125 laser melting printer. Figure 4As shown in the diagram. In the first component, the body-centered cubic lattice structure, the support column length is 8mm, the column diameter is 0.5mm, and the included angle between adjacent supports is 45°. In the second component, the support column length is 8mm, the column diameter is 2mm, and the included angle between adjacent supports is 45°. In the third component, the diamond-like carbon lattice structure, the support column length is 1mm, the column radius is 0.2mm, and the included angle between adjacent supports is 109.5°. Manufacturing defects, including bending, protrusion, breakage, missing supports, and excessively thick supports, were artificially introduced during component fabrication to simulate defects that might exist in actual engineering applications.

[0045] In this embodiment, to determine the parameters of the present invention, three 316L stainless steel lattice structure components were inspected using CT equipment to obtain a sufficient number of CT image datasets for training and evaluating the method model. Two types of CT equipment were used for data acquisition in this embodiment to enhance the model's generalization ability to differences in CT images captured by different scanning devices. A dataset of 4636 CT slices, including body-centered cubic and diamond-like carbon lattice structures, was collected. The data was divided into training and test sets in an 8:2 ratio for training and evaluating the model, respectively. Defect labels were added by professionals. To prevent overfitting, conventional data enhancement processes were performed: cropping, flipping, and channel conversion.

[0046] In this embodiment, a single NVIDIA GeForce RTX 3090 (24GB) GPU and a 13th-generation Intel Core i9-13900K were used to train all models. The total epochs were 50, and other parameters were set to default. Training was conducted using the deep learning framework PyTorch.

[0047] In this embodiment, to demonstrate the superiority of the present invention, it is compared with the state-of-the-art object recognition method—YOLOv10 algorithm. The effectiveness of the model is evaluated using metrics such as precision (P), recall (R), and average precision (mAP). Precision (P) represents the number of expected positive samples, recall (R) represents the number of expected accurate positive samples in the sample, the area under the precision-recall curve is AP, and the average AP value for each object is called mAP.

[0048] like Figure 1 The intelligent defect detection method for CT images of additively manufactured multi-type lattice structural components, as shown, includes the following steps:

[0049] S1. A backbone network composed of several convolutional normalized activation functions and pooling functions with different parameters is used to process the original CT images of various types of lattice structure components manufactured by additive manufacturing, extract the features of the input images, and generate feature maps of multiple scales; step S1 includes the following steps:

[0050] S1-1. Acquire original CT images of various types of lattice structure components manufactured by additive manufacturing, adjust the adaptive sensitivity field through deformable convolutional layers, and simultaneously perceive regular or irregular shapes to improve feature representation, generate feature maps with rich spatial information, and obtain the original features.

[0051] S1-2. Perform three max pooling operations on the feature map obtained in step S1-1. The size of all pooling kernels is fixed at 5×5 and the stride size is 1. Extract global multi-scale information while keeping the feature size unchanged to obtain pooled features.

[0052] S1-3. The original features obtained in step S1-1 and the pooled features obtained in step S1-2 are concatenated along the channel dimension, and the channels are compressed through a convolutional layer to output the fused feature map; S2. A path aggregation network consisting of up-convolution, upsampling and downsampling operations is used as the neck network to fuse the feature maps of different scales generated in step S1 together to generate a feature map with multi-scale information, thereby improving the detection accuracy of targets at different scales;

[0053] like Figure 2 As shown, the SPPDF module in the backbone network adds deformable convolutions to the first convolution based on the SPPF (Spatial Pyramid Pooling Fast) module, changing the sampling position in a certain way so that it can capture target defects with a series of shapes, sizes, and deformations. The core idea of ​​the SPPF module is to perform multi-scale processing on the input feature map through multiple max pooling operations with the same pooling kernel size, thereby enhancing the model's receptive field and its ability to perceive objects at different scales.

[0054] S2. A path aggregation network consisting of upconvolution, upsampling, and downsampling operations is used as the neck network to fuse the feature maps of different scales generated in step S1, generating a feature map with multi-scale information, thereby improving the detection accuracy of targets at different scales; step S2 includes the following steps:

[0055] S2-1. Encode the position of the input feature map along the width W and height H directions, concatenate the position codes with the feature mapping directions of the two adjacent layers, and feed them into different 1×1 convolutions respectively;

[0056] S2-2. Use a convolutional block consisting of the SiLU activation function, batch normalization, and convolution operation to obtain the channel dimension information C;

[0057] S2-3. Use the Sigmoid activation function and convolution operation to weight the feature map, retain the shallow spatial details and incorporate the deep semantic information to obtain a feature map that balances localization accuracy and semantic expression.

[0058] S3. In the head network, the feature map with multi-scale information generated by the neck network in step S2 is compared with the labeled image data, and the loss function of this training set is calculated. The loss function includes localization loss and classification loss; the location information, category and confidence of the defect in the image are output to complete the intelligent defect detection of CT images of additive manufacturing multi-type dot matrix structural parts; step S3 includes the following steps:

[0059] S3-1. Use a multi-task joint optimization loss function to train the model:

[0060] L=λ box L CIoU +λ dfl L DFL +λ cls L BCE (1)

[0061] In equation (1), L CIoU For the complete intersection-union bounding box loss function; L DFL L is the distribution focus loss function; BCE Let λ be the classification loss function; box , λ dfl , λ cls Both are weights of the subtask loss function, and λ box =7.5, λ dfl =1.5, λ cls =0.5;

[0062] S3-2. For each sample, define the true label g = [x g ,y g ,w g ,h g ,c g The prediction result is p = [x] p ,y p ,w p ,h p ,c p ], then L CIoU The loss function for a perfect intersection-union bounding box is:

[0063]

[0064] In equation (2), ρ 2 =(x g -x p ) 2 +(y g -y p ) 2 ;

[0065] IoU is the intersection-union ratio of the predicted bounding box and the ground truth bounding box; c is the diagonal length of the bounding rectangle of the two boxes;

[0066] S3-3. Let the boundary coordinates of the true label of the target to be detected be t. Discretize the possible values ​​of the boundary coordinates into 16 intervals, and use the network to predict the probability of each interval to obtain S. i and S i+1 L DFL The distributed focus loss function is:

[0067] L DFL =-((t) i+1 -t)log(S i )+(tt i )log(S i+1 (3) S3-4, Classification loss function L BCE for:

[0068] L BCE =-(g c log(p c )+(1-g c log(1-p) c (4)

[0069] In equation (4), g c p represents the label value for the category. c The confidence level of the class predicted by the network;

[0070] S3-5. The feature map is decoded through steps S3-1 to S3-4, and the location information, category and confidence level of the defect in the image are output.

[0071] The comparison results between this embodiment and the YOLOv10 target recognition algorithm are shown below.

[0072] I. Comparison of Training Loss Values

[0073] like Figure 5 As shown, both this embodiment and the YOLOv10 target recognition algorithm were trained for 50 epochs, and the loss values ​​were very close, so it can be considered that the convergence speed of the two models is the same.

[0074] II. Comparison of Precision-Recall Curves for Various Types of Defects

[0075] like Figure 6 As shown in Table 1, the detection accuracy of this embodiment and the YOLOv10 target recognition algorithm for various defects is as follows.

[0076] Table 1 shows the detection accuracy of this embodiment and the YOLO v10 target recognition algorithm for various defects.

[0077]

[0078] As shown in Table 1, this embodiment improved the detected mAP by about 4%, reaching 97.19%. The detection accuracy for bending, protrusion, missing, and excessively thick supports was also improved. Furthermore, the detection accuracy for breakpoints was increased from 85.04% to 93.53%, an improvement of 8.31%. The above results demonstrate the effectiveness and high accuracy of the present invention.

[0079] III. Comparison of CT slice images containing different types of defects

[0080] like Figure 7 As shown, the YOLO v10 algorithm can only automatically identify two excessively thick pillars and one breakpoint, while this embodiment can detect all defects (two excessively thick pillars and two breakpoints).

[0081] like Figure 8 As shown, the YOLO v10 algorithm can only automatically identify three prominent defects, while this embodiment can detect four prominent defects.

[0082] like Figure 9 As shown, the YOLO v10 algorithm can only automatically identify three breakpoint defects, while this embodiment can detect defects with four breakpoints.

[0083] In summary, this invention, based on the YOLO v10 algorithm, introduces a fast deformable spatial pyramid pooling module and an attention mechanism module. The fast deformable spatial pyramid pooling layer is used to improve multi-scale defects and spatial information, while the attention mechanism module is mainly used to improve the feature details of defects. This invention can obtain more spatial detail information and global information of target defects, outputting more expressive feature maps. The accuracy of automated defect identification is significantly better than that of the YOLO v10 model, ensuring high-efficiency and high-accuracy detection of defects (bending, protrusion, fracture, missing support, and excessively thick support, etc.) within various types of lattice structures.

[0084] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for intelligent defect detection of additively manufactured multi-type dot matrix structural components based on CT images, characterized in that, Includes the following steps: S1. A backbone network composed of several convolutional normalized activation functions and pooling functions with different parameters is used to process the original CT images of additive manufacturing multi-type lattice structure components, extract the features of the input images, and generate multiple feature maps of different scales. S2. A path aggregation network consisting of upconvolution, upsampling and downsampling operations is used as the neck network to fuse the feature maps of different scales generated in step S1, thereby generating a feature map with multi-scale information and improving the detection accuracy of targets at different scales. S3. In the head network, the feature map with multi-scale information generated by the neck network in step S2 is compared with the labeled image data, and the loss function of this training set is calculated. The loss function includes localization loss and classification loss. It outputs the location information, category, and confidence level of defects in the image, and completes intelligent defect detection of CT images of various types of additive manufacturing lattice structural parts.

2. The intelligent defect detection method for CT images of additively manufactured multi-type dot matrix structural components according to claim 1, characterized in that: Step S1 includes the following steps: S1-1. Acquire original CT images of various types of lattice structure components manufactured by additive manufacturing, adjust the adaptive sensitivity field through deformable convolutional layers, and simultaneously perceive regular or irregular shapes to improve feature representation, generate feature maps with rich spatial information, and obtain the original features. S1-2. Perform three max pooling operations on the feature map obtained in step S1-1. The size of all pooling kernels is fixed at 5×5 and the stride size is 1. Extract global multi-scale information while keeping the feature size unchanged to obtain pooled features. S1-3. The original features obtained in step S1-1 and the pooled features obtained in step S1-2 are concatenated along the channel dimension, and the channels are compressed through a convolutional layer to output the fused feature map.

3. The intelligent defect detection method for CT images of additively manufactured multi-type dot matrix structural parts according to claim 1, characterized in that: Step S2 includes the following steps: S2-1. Encode the position of the input feature map along the width W and height H directions, concatenate the position codes with the feature mapping directions of the two adjacent layers, and feed them into different 1×1 convolutions respectively; S2-2. Use a convolutional block consisting of the SiLU activation function, batch normalization, and convolution operation to obtain the channel dimension information C; S2-3. Use the Sigmoid activation function and convolution operation to weight the feature map, retain the shallow spatial details and incorporate the deep semantic information to obtain a feature map that balances localization accuracy and semantic expression.

4. The intelligent defect detection method for CT images of additively manufactured multi-type dot matrix structural components according to claim 1, characterized in that: Step S3 includes the following steps: S3-1. Use a multi-task joint optimization loss function to train the model: L=λ box L CIoU +λ dfl L DFL +λ cls L BCE (1) In equation (1), L CIoU For the complete intersection-union bounding box loss function; L DFL L is the distribution focus loss function; BCE Let λ be the classification loss function; box , λ dfl , λ cls Both are weights of the subtask loss function, and λ box =7.5, λ dfl =1.5, λ cls =0.5; S3-2. For each sample, define the true label g = [x g ,y g ,w g ,h g ,c g The prediction result is p = [x] p ,y p ,w p ,h p ,c p ], then L CIoU The loss function for a perfect intersection-union bounding box is: In equation (2), ρ 2 =(x g -x p ) 2 +(y g -y p ) 2 ; IoU is the intersection-union ratio of the predicted bounding box and the ground truth bounding box; c is the diagonal length of the bounding rectangle of the two boxes; S3-3. Let the boundary coordinates of the true label of the target to be detected be t. Discretize the possible values ​​of the boundary coordinates into 16 intervals, and use the network to predict the probability of each interval to obtain S. i and S i+1 L DFL The distributed focus loss function is: L DFL =-((t i+1 -t)log(S i )+(t-t i )log(S i+1 )); (3) S3-4, Classification Loss Function L BCE for: L BCE =-(g c log(p c )+(1-g c )log(1-p c )); (4) In equation (4), g c p represents the label value for the category. c The confidence level of the class predicted by the network; S3-5. The feature map is decoded through steps S3-1 to S3-4, and the location information, category and confidence level of the defect in the image are output.

Images